Adaptive Folk

The Dream

I have always wanted to see 'smart' NPCs that use AI to become dynamic entities that respond realistically and with memory. This is my first attempt at making that dream a reality.

Important Note: AI has limitations and dangers that I am aware of, but having AI-powered NPCs is something that I believe can create unique, immersive, and happy experiences for players. That has been a powerful enough motivation for me to continue creating this mod.

---

Overview

Kweebec are transformed from static NPCs into AI-powered interactive characters.

Each Kweebec:

• Has a unique name
• Responds in-character using locally-generated AI dialogue
• Remembers past conversations

The result is a more immersive world where Kweebec feel more aware, reactive, and alive.

For Example:

• I asked a Kweebec who Harry Potter was and it did not know
• I taught the Kweebec who Harry Potter was
• I asked the Kweebec about Harry Potter and it was able to tell me about him

---

Features

NPC Identity

• Name assigned to NPC on spawn from a curated list of 'plant-people-like' names

AI-powered Dialogue

• AI-powered conversations with any Kweebec via in-game chat
• Fully local via Ollama backend (Open-source software)
  • No costs
  • No limits
  • No internet required
• Low-latency (typically under 2 seconds)

Persistent Memory System

• NPC conversation history stored in a local document (Path to AI memory stored here: (world)/mods/Adaptive_Folk_Kweebecs)
  • Summarized and used by AI to respond in-character as a Kweebec
• NPCs can reference previous interactions with multiple players

Graceful Fallback Mode

• If AI is not setup (Ollama + any AI LLM) Kweebec still retain their new name and respond using simple fallback responses.

---

Getting Started

Full Experience (Names, AI-Powered Kweebec Responses, Conversation Memory)

• Place .jar in Hytale 'Mods' folder
• Install and run Ollama Open-source software (https://ollama.com/download). This is the software that connects the mod to your local AI
• Use any CLI (powershell or command line on windows) to pull a model for Ollama to use using this command ollama pull llama3.2:3b
  • (Optional) Substitute llama3.2:3b with ANY available model (https://ollama.com/library)
• Launch Hytale, select Adaptive Folk mod, and speak to any Kweebec!

Minimal Experience (Names, Basic Kweebec Responses)

• Place .jar in Hytale 'Mods' folder
• Launch Hytale, select Adaptive Folk mod, and speak to any Kweebec!

---

Additional Information Regarding LLMs

Recommended list of LLMs:

• llama3.2:3b (My recommendation AND my preferred model while testing on my RTX 3060 Ti)
• llama3.2:1b (My recommendation for computers with less power than an RTX 3060 Ti)
• gemma3:4b
• gemma3:1b
• qwen2.5:3b
• qwen2.5:7b
• mistral:7b
• gemma3:12b

Important Notes about local LLM use:

• Any LLM from the Ollama library above will work IF your hardware supports it. I found my RTX 3060 Ti (8gb VRAM) to run very well with the llama3.2:3b model listed above. VRAM will likely be your bottleneck here
• Guardrails vary from model to model! I found llama3.2:3b to be a very good model for safe responses

Ollama Command Examples

• Download model: ollama pull llama3.2:3b
• Remove model: ollama rm llama3.2:3b
• View all models: ollama list

---

Special Thanks:

• Ollama
• https://britakee-studios.gitbook.io/hytale-modding-documentation (I forked off of this template provided at this site to get started: https://github.com/realBritakee/hytale-template-plugin)
• https://hytalemodding.dev/en/docs
• https://hytale.limetta.dev/
• Scheletro

---

FAQ:

Can Kweebec remember interactions with multiple players?

Yes. Kweebec can reference interactions across different players.

Which AI models are supported?

Any model supported by Ollama should work, as long as your hardware can handle it.

What affects performance?

Performance depends primarily on:

• Your CPU/GPU
• Available VRAM
• The size of the model you are running

Smaller models generally respond faster. Pay attention to the 'b' size. The higher the 'b' the more powerful and resource intensive the model typically is (e.g. '1b' is much more performant than a '3b' model).

Why is a Kweebec not responding with AI?

• Ollama is not installed and running
• No model is installed

In these cases, basic fallback responses will be used.

Is this safe to use?

• Ollama is Open-source
• All AI runs locally on your machine. No data is sent externally.
• The AI models I tested have had good guardrails in place BUT there is always some risk of AI responding poorly.